The release of DeepSeek’s R1 model at the end of January caused a selloff in stocks exposed to AI infrastructure. In part, this was due to it introducing innovations that showcased impressive reductions in the cost of inferencing under certain conditions – doing more with less, it seemed.
We believe that further reducing the cost of inference, the process of generating ideas or reasoning from AI models, acts as a substantial tailwind for the continued advancement of AI and for the beneficiaries of AI adoption.
Falling cost benefits
It should be noted that the collapsing cost of inference is not a new dynamic and is a key characteristic of technology progression. The cost of inference (measured in dollars per million tokens) had fallen 20x even pre-DeepSeek. Indeed, the open-source nature of Meta’s Llama models, from which R1 itself was party derived, is intended to create these types of innovation at the trailing edge, refining and advancing the leading-edge work.
A falling cost of inference benefits the continued scaling of AI, through cheaper test-time compute as well as adoption through lower cost inputs when running AI inference and implementing it within applications.
Dramatically more capable models at the same cost will speed up innovation or, put another way, the cost of running the current most advanced models will collapse over time. For example, the initial cost for querying OpenAI’s most capable ‘o3’ model at the level that delivered breakthrough performance on the ARC-AGI test, one designed to measure progress towards artificial general intelligence (AGI), was near $3,000 at launch. However, OpenAI API (application programming interface) costs decreased 90% in 2024 and, per recent comments from CEO Sam Altman, are expected to decrease the same amount again this year. While they are not direct corollaries, it highlights the likely magnitude of continued cost reduction over time.
This raises the possibility of introducing beyond-PhD level expertise into workflows at minimal incremental cost. It is feasible to imagine real-time AI assistance across the full range of medical expertise on Intuitive Surgical’s robots, or a researcher having access to cross-disciplinary scientific expertise in real time as they conduct experiments. Low-cost inferencing of more advanced AI will further the ability of companies such as Walmart, Tesco and Dick’s Sporting Goods to solve the most complex inventory management and logistics problems in ways their peers are already unable to match.
The cost coming down means you can drive more ubiquity of what were features that once were sort of premium tier. - CEO Satya Nadella, Microsoft earnings call 4Q24
Importantly, we are seeing a continued decline in the cost of inference for all models, not just text-based models such as DeepSeek. It is the greater accessibility of the most capable and the multimodal models, those that can input and output across text, video and speech, that is so exciting. These are the models that offer the greatest utility to companies and we believe the progression to AGI will be led by multimodal models.
Broader AI adoption
While cost is one consideration in a company’s adoption of AI in the short term, we believe that a lower cost of inferencing ultimately acts as a significant tailwind to both future AI development and the democratisation and widespread adoption of AI at a corporate level.
It is paramount to identify companies that are best positioned to capitalise on this falling cost and able to broaden their implementation of AI across their business. We have already seen companies adopt AI to gain a competitive edge, such as London Stock Exchange Group and Publicis where in both cases the use of AI in their products and processes has allowed them to substantially outgrow their peers over the past year and take meaningful market share. It is the ability to capture both outsized profit growth and command premium market valuations that we believe offers the potential for significant investment returns over the long term.
There already exists a significant and growing divergence in the ability of companies to react to these technology developments and capture the disruptive economics of AI rather than suffer significant value destruction. The recent developments will only exacerbate this disparity between the winners and losers, particularly so beyond the traditional technology sector.
The pace of corporate adoption of AI has positively surprised us over the past two years as companies have identified the opportunity the technology presents. We believe identifying companies on the right side of the disruption AI will bring to all sectors is one of the most important investment considerations of the coming years.
It will not be long before companies adopting this technology [AI] start to outperform their peers, and it will not be easy to tell why from the outside. - Sam Altman, CEO OpenAI
The DeepSeek announcement is but one component of the innovation that is so compelling. We have also heard commitments to the latest generation of frontier models from major players including Meta Platforms and OpenAI, with the expectation that this next generation will bring the same advancement as the current one did in terms of capabilities.
We remain incredibly constructive on the outlook for companies that stand to benefit from the adoption of AI. Recent developments have only enhanced our conviction that the democratisation of leading-edge AI will disrupt vast swathes of the global economy and introduce incredible investment opportunities.
Combining our knowledge of the underlying technology developments, our large team’s ability to operate across all parts of the AI ecosystem, and our experience of just how rapidly and meaningfully disruption can occur with markets, is core to our approach for investing across all sectors in this new-look world. It is this opportunity that the Polar Capital Artificial Intelligence Fund was launched over seven years ago to capture and our excitement at what lies ahead is greater than ever.
